It is well known, that the expectation and the essential supremum qualify as time consistent (deterministic) risk functionals. We show that if one allows risk functionals at random level, the value function corresponding to the optimal decisions evolves as a martingale and a dynamic programming principle is applicable. A verification theorem allows to characterize optimal decision by sub- and supermartingales for these seemingly time-inconsistent models.  We also show how to decompose a risk functional into extended conditional risk functionals in such a way that the original functional is regained by compounding the conditional functionals in an appropriate manner thus ensuring time-consistency in an extended sense. The way to achieve this result is through change-of-measures and duality. We also show that without the change-of-measures approach the only time consistent risk functionals having a Kusuoka representation are the expectation and the essential supremum.